Classification is one of the most important tasks in machine learning. Conventional classification methods aim to attain low recognition error rate and assume the same loss from different kinds of misclassifications. However, in the applications such as the doorlocker system based on face recognition, software defect prediction and multi-label learning, different kinds of misclassification will lead to different losses. This requires the learning methods to pay more attention to the samples with high-cost misclassification, and thus make the total misclassification losses minimized. To deal with this issue, cost-sensitive learning has received the considerable attention from the researchers. This study takes the theoretical foundation of cost-sensitive learning as the focal point to analyze and survey its main models and the typical applications. At last, the difficulty and probable development trend of cost-sensitive learning are discussed.